This approximation seems the same proposed in Sønderby et al 2016

[silt99]

The use of a GAN in the null space for superresolution have already been done, first in C. Sonderby, J. Caballero, L. Theis, W. Shi, and F. Huszar, “Amortised MAP inference for image super-resolution,” in International Conference on Learning Representations, 2017 (https://arxiv.org/abs/1610.04490). The only difference that I see between the two approaches is that they use a CNN to approximate the pseudo-inverse. Is this accurate?

[wyhuai]

Thanks for your question! I did not notice this paper but it is interesting. There are many differences in details. For example, we propose a special sampling operator rather than using a CNN for approximation. Also, the training loss functions are different. Besides, we had a follow-up work published in ICLR 2023 maybe you are interested.